HawkFranklin-Research/UAP-May8-Embeddings

UAP May 8 Embeddings

This dataset is a derived analysis package built from the public May 8 release of U.S. government UAP/UFO files.

It is meant for downstream search, clustering, anomaly review, and retrieval experiments. The raw archive is published separately; this repo contains the analysis outputs and vector representations.

Source

Primary archive:

• —HawkFranklin-Research/UAP-May8

Local analysis workspace used to generate this dataset:

• —analysis/

Contents

This repo contains two embedding views:

• —documents.csv and document_embeddings.parquet
• —chunks.csv and chunk_embeddings.parquet

It also includes:

• —document_embeddings.npz
• —chunk_embeddings.npz
• —REPORT.md
• —inventory_summary.json
• —clustering outputs
• —anomaly outputs
• —figures

Dataset Shape

What Was Removed For Public Release

The public tables keep the content and analysis fields, but they drop operational fields that are only useful inside the local workspace, including:

• —absolute local file paths
• —key-tracking fields
• —attempt bookkeeping

That keeps the published dataset self-contained and avoids leaking local environment details.

Recommended Files

For most use cases, start with:

• —document_embeddings.parquet
• —chunk_embeddings.parquet
• —REPORT.md

Use the .csv files if you want quick inspection without loading parquet. Use the .npz files if you want direct NumPy array access.

Columns

Document embeddings

The document table includes:

• —document identity
• —source agency
• —file group
• —release and incident metadata
• —file size and media characteristics
• —the 3072-dimensional embedding vector

Chunk embeddings

The chunk table includes:

• —document identity
• —chunk identity
• —page ranges or sampled-frame metadata
• —media type
• —chunk text metadata
• —the 3072-dimensional embedding vector

Loading

Document-level example:

from datasets import load_dataset

docs = load_dataset(
    "HawkFranklin-Research/UAP-May8-Embeddings",
    data_files="document_embeddings.parquet",
    split="train",
)

Chunk-level example:

from datasets import load_dataset

chunks = load_dataset(
    "HawkFranklin-Research/UAP-May8-Embeddings",
    data_files="chunk_embeddings.parquet",
    split="train",
)

Notes

• —This is a derived representation of the archive, not the original evidence.
• —The embeddings are suitable for similarity search, clustering, and anomaly detection.
• —The dataset should be interpreted alongside the original archive and report, not instead of them.